Matching of signal, noise and hardware timescales for filtering and forecasting of correlated noise signals

arXiv:2610.10037 · cs.LG, cond-mat.dis-nn, physics.app-ph, physics.data-an · Submitted 2026-10-07 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Matching of signal, noise and hardware timescales for filtering and forecasting of correlated noise signals".

Jane: The gist: Timescale matching between noise correlation time, reservoir memory, and forecast horizon determines whether correlated noise is filtered or predicted by a physical reservoir system.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’re looking at this paper, "Matching of signal, noise and hardware timescales for filtering and forecasting of correlated noise signals." It’s really about how the timing of the input data matches the timing built into the physical hardware itself.

Jane: Exactly. It suggests that whether a physical system filters out noise or tries to predict it depends entirely on how long that noise lasts compared to how much memory the reservoir has and how far out you want to predict.

Lu: The core idea here is that we need to match these timescales: the correlation time of the noise, the memory timescale of the reservoir, and then your forecast horizon. It’s about finding that sweet spot where things either get averaged out or where you can actually make a prediction.

Meng: From a deployment standpoint, it sounds like they are trying to figure out when to use a physical system for filtering versus when an algorithm should take over because the hardware just isn't built for that specific noise timing.

Lalam: I see it as defining the operating window for temporal inference based on those three interacting timescales, which is a crucial concept for any real-world signal processing task.

Tom: The paper tests this using a physical reservoir made from a nanoporous niobium oxide layer with volatile switching dynamics, and they mix deterministic signals with colored noise to see how it performs.

Jane: They set up this setup where they can control the noise correlation time, which they call tau zero ranging from six point three times ten-seven seconds up to one thousand times that value, and they look at how that affects their results.

Lu: The paper introduces a Forecasting Regime Index, or FRI, defined as FRI = ten

tau / sqrt tau m tau w: , where tau is the noise correlation time, tau m is the reservoir memory horizon, and tau w is the forecast window size <ref:2610.10037#pg2>.

Tom: If that FRI is less than zero, it means the noise evolves faster than both the prediction window and the memory timescale, which points toward reservoir-based averaging or some other kind of forecasting.

Title and authors: Jane: But if that correlation time is longer than both those timescales, then they find that the physical reservoir treats that colored noise as part of the signal itself, which lets them reconstruct and predict it.

Meng: That distinction is important because it tells you whether you should trust the hardware to smooth out fast noise or whether you need a separate algorithm to handle slower correlated structures.

Lalam: The results show that for reconstructing and predicting signals with noise injected into them, the normalized root mean square error, or NRMSE values, are lowest when the noise correlation times are longer.

Tom: And they test this on real-world data involving four cryptocurrencies: Bitcoin, Ethereum, Solana, and Dogecoin volatility. The finding is that NRMSE values decrease as the correlation time increases.

Jane: That’s interesting because it confirms what they saw with the synthetic noise signals for longer correlation times, which tells us that slower noise structures are actually more useful for forecasting in these markets.

Lu: They pinpoint specific optimal conditions, showing that the lowest NRMSE values occur when the forecast window size is smaller than fifteen point eight seven tau zero but also when the correlation times are larger than that future window size.

Tom: So, what this means for us is that we can start designing physical reservoir architectures based on these timescale matches instead of just picking a standard hardware setup and hoping it works.

Jane: It suggests we need to look at the autocorrelation structure and spectral content of our input data to figure out the best way to map that signal into the memory timescales where the reservoir is actually effective.

Lu: The paper argues that coloured noise isn't just a nuisance; it’s a structured component of what you are trying to model, and successful prediction depends on encoding that structure into those specific hardware memory times.

Meng: If we are shipping models to production, this implies that our input encoding shouldn't just be based on the raw sampling interval but should be tailored to match the physical constraints of the system we’re using.

Title and authors: Lalam: The fading memory of the reservoir essentially defines a practical operating window for temporal inference when dealing with noisy data.

Tom: I think this shifts how we approach model design, suggesting that optimal encoding should be determined by autocorrelation structure and spectral content rather than just raw sampling interval.

Jane: And it also suggests that the reservoir outputs might contain richer information than just a direct regression because it mixes timescales of the input data with its own dynamics and filters things out.

Lu: The paper demonstrates these design principles for encoding noisy temporal data for physical hardware and developing novel physical reservoirs based on these timescale matches.

Meng: For deployment, this means we need to consider the hardware constraints when we select our AI architecture, especially when dealing with signals that have multiple overlapping time scales.

Lalam: It gives us a clear framework: if the noise is fast, it gets averaged by the reservoir; if it’s slower and correlated with the forecast horizon, you can actually forecast that component separately.

Tom: So to wrap up on "Matching of signal, noise and hardware timescales for filtering and forecasting of correlated noise signals," this work shows that we need to consider the relationship between the input signal's correlation time, the reservoir’s memory capacity, and your prediction goal to decide if you should filter or predict.

Jane: It really emphasizes that in physical systems like these reservoirs, what works depends on matching those specific temporal structures rather than just brute-forcing a solution.

Lu: The implication is that optimal encoding needs to be determined by the autocorrelation structure and spectral content of the signal, not just raw sampling intervals.

Meng: And for us at the startup side, it means we need to be mindful of these hardware timescales when selecting our AI architecture for deployment on real-world data.

Lalam: It gives us a way to handle complex signals where deterministic and stochastic components are mixed together by treating the noise structure as part of the learning problem.

The paper's summary: Tom: So, to wrap up that first part, this paper is about finding that precise match between three different timescales: how fast the noise is changing, how much memory the physical reservoir has, and how far out you want to predict a signal.

Jane: Right. And it’s not just some abstract theory; they built a physical system—a type of oxide film with some defects acting like a memristor—and they used colored noise mixed with deterministic signals to test this idea.

Lu: The core finding is that the performance of the reservoir really hinges on whether the noise correlation time is shorter or longer than both the reservoir's memory range and your forecast window size.

Tom: If those timescales match up right, say they are long enough, then you get better results by treating that correlated noise as part of the signal itself rather than just trying to filter it out.

Jane: That means if you’re dealing with slow, structured noise, the physical system can actually help predict it instead of just averaging it down into oblivion.

Lu: They introduced this Forecasting Regime Index, or FRI—it’s a calculation using logarithms involving the noise correlation time over the memory and window sizes—to tell you whether you are in one of two regimes.

Tom: If that index is less than zero, it tells you the reservoir is just doing averaging, which isn't as good for prediction tasks.

Jane: But if that index is greater than zero, then the reservoir can actually reconstruct and predict the signal when it’s longer than those timescales.

Lu: They showed that for reconstructing signals with noise injected into them, the normalized root mean square error, or NRMSE values—the measure of how wrong you are—gets lower as that noise correlation time gets longer.

Tom: And they took this from a lab setup to real-world cryptocurrency data, testing Bitcoin and Ethereum volatility.

Jane: The results on those crypto signals showed that the error rates dropped when the correlation time increased, which lines up with what they found in their synthetic noise tests for longer correlation times.

Lu: Specifically, they found the best predictions happen when the future window size is smaller than about fifteen point eight seven times tau zero but only if the noise correlation time is larger than that forecast window size.

Tom: So, what this means for a producer of AI models or a researcher? It suggests that we shouldn't just assume raw sampling intervals are the only thing that matters when we design encoding.

Jane: We need to look at the autocorrelation structure and the spectral content of our input data—what the noise looks like—to determine how to map that information into the memory timescales where our AI is actually effective.

Lu: The paper argues that colored noise isn't just some random nuisance; it’s a structured component of what you are trying to model, and successful prediction depends on encoding that structure into those specific hardware memory times.

Tom: So, the main conceptual move here is treating the noise as a signal feature itself.

Jane: And they show that this fading memory of the reservoir defines a practical operating window for temporal inference when you have noisy data.

Lu: The implication is that optimal encoding needs to be determined by the autocorrelation structure and spectral content of the signal, not just raw sampling intervals.

Tom: And it also suggests that the reservoir outputs might contain richer information than just a direct regression because it mixes timescales of input data with its own dynamics and filters things out.

Jane: So, for anyone using AI for forecasting on noisy time series, this paper gives you a concrete way to decide if you should be filtering out the noise or trying to predict the structure it’s hiding.

Lu: It’s a practical guide for designing physical reservoir architectures based on these timescale matches instead of just picking a standard hardware setup and hoping it works.

Tom: Right, so next time you're designing a physical system, check those correlation times against the memory limits before you commit to an architecture.

The paper's improvements: Tom: So, we just looked at how timescale matching dictates whether you filter or predict noise, and now let’s talk about what they suggest we *do* differently with that knowledge.

Jane: The paper points out that the real improvement isn't just running the experiment; it’s changing how we design the input encoding process itself.

Lu: They argue for designing an optimal encoding process that maps the informative temporal structure of a signal, including its noise correlations, into memory timescales where the reservoir is actually computationally effective.

Tom: That means we stop relying only on raw sampling intervals to define our data points; instead, we need to use the autocorrelation structure and spectral content to guide the mapping.

Jane: It suggests that we should be building reservoir architectures specifically guided by these timescale matches, not just picking some random hardware configuration.

Lu: That’s a big idea—it means the architecture should be tailored based on what you are trying to filter or predict across those different temporal scales of noise.

Tom: And they also hint that the reservoir outputs themselves might carry richer information than just a simple regression because it mixes input data timescales with its own internal dynamics and filtering.

Jane: It means we can’t just treat the output as a clean prediction; it’s actually a blend of what was in the input and how long the reservoir's memory lasts.

Lu: This opens up new avenues for how we interpret those outputs, moving beyond direct regression to look at those mixed timescales.

Tom: So, if we can get that encoding right—matching the noise structure to the hardware memory—we might actually unlock better performance than what a standard setup could achieve.

Jane: And this leads into the next part of their work where they show how this principle applies directly to real-world financial markets, specifically cryptocurrency volatility.

Lu: They found that for those crypto signals, NRMSE values get lower as the correlation time increases, which is consistent with what they saw in the synthetic noise tests for longer correlation times.

Tom: So, we’re seeing a pattern: longer correlated noise structures help improve forecasting accuracy in these real-world applications.

Jane: And they give us a specific sweet spot—the best performance happens when the future window size is smaller than fifteen point eight seven times tau zero, but only if the noise correlation time is larger than that future window size.

Lu: That’s a very specific constraint for when you apply this theory to practical forecasting problems.

Tom: So, to summarize the improvements, it's not just about bigger hardware; it’s about smarter input preparation and architecture design based on those three interacting timescales.

Jane: It really pushes us toward designing systems that are inherently aware of their temporal relationships with the data they’re processing.

Lu: And this framework provides a way to handle complex signals where deterministic and stochastic components are mixed together by treating the noise structure as part of the learning problem.

Tom: If we can implement these encoding strategies effectively, we might see better generalization when deploying these reservoir systems on real traffic, though they don't really show what happens at massive scale yet.

Conclusion: Tom: So we’re wrapping up on "Matching of signal, noise and hardware timescales for filtering and forecasting of correlated noise signals." Basically, this paper shows that performance depends entirely on matching the noise correlation time to the reservoir's memory and your forecast horizon.

Jane: Right. The big implication is moving beyond just building a generic AI model; you have to consider the physical constraints of your hardware when you design how the input signal is encoded.

Lu: They’ve established this idea that colored noise isn't just some random nuisance; it's a structured component that needs to be handled by matching those specific temporal scales.

Tom: So, if we want better results, we need to determine the optimal encoding based on autocorrelation structure and spectral content instead of just how fast you’re sampling the data.

Jane: That’s a huge shift for anyone building models for real-world time series data; it’s about tailoring the input specifically to what the hardware can actually compute efficiently.

Lu: It suggests that successful prediction is governed by that timescale matching between the input and the hardware, which is a more practical operating window for temporal inference.

Tom: And they showed this principle works on real cryptocurrency volatility data, confirming that longer correlation times lead to lower error rates in those financial predictions.

Jane: It gives us a concrete rule of thumb: for these types of complex signals, if you want good results, try to ensure the noise correlation time is longer than your forecast window size.

Lu: That’s what the Forecasting Regime Index was built for—it automatically tells you whether you should be filtering or predicting based on those three timescales.

Tom: It’s a solid piece of engineering because it gives us a clear metric to judge if we're in the right regime.

Jane: So, this paper moves the conversation from just "how to train an AI" to "how do we design the hardware and input pipeline so the AI can actually do its job efficiently?"

Lu: It’s a strong foundation for future work, suggesting that understanding these physical constraints is a necessary step before we can build more advanced systems.

Tom: Exactly, and I'm looking forward to seeing how this timescale matching concept applies to other areas of physical system modeling.

Jane: Definitely. And next week, we’re diving into something completely different—we’ll be looking at how memorization is encoded in the geometry of diffusion models before they even generate a single output.

Lu: That sounds interesting; I think linking those geometric features to auditing model outputs could reveal a lot about why some models produce perfect copies and others don't.

Meng: From an engineering standpoint, I’m curious how much overhead this geometric probing adds to the training time compared to just looking at the final sample.

Tom: We’ll get to that, Meng. But for now, keep your eyes peeled for our next paper on diffusion models and their energy landscapes.

Joshua Donald, Alex Gabbitas, Arthur G. T. Coveney, Sergey Savel’ev, Pavel Borisov

Department of Physics, Loughborough University

cs.LG, cond-mat.dis-nn, physics.app-ph, physics.data-an

Submitted: 2026-10-07

Updated: 2026-10-07

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: The gist: Timescale matching between noise correlation time, reservoir memory, and forecast horizon determines whether correlated noise is filtered or predicted by a physical reservoir system.

Key concepts

Physical Reservoir Computing (RC): This technique uses the nonlinear dynamics of physical systems, like a specific oxide film, to process time-dependent data efficiently. It exploits fixed intrinsic response timescales in hardware to handle real-world signals that have both deterministic and stochastic components across various time scales.
Timescale Matching: This is the core hypothesis: performance depends on aligning three timescales—noise correlation time, reservoir memory, and forecast horizon. If these are mismatched, the system either averages out noise or incorporates it into predictions. Optimal prediction occurs when these timescales align appropriately.
Forecasting Regime Index (FRI): An empirically derived index (FRI=log10[τ/√τmτw]) used to classify the noise behavior. An FRI < 0 suggests the noise evolves faster than both key timescales, indicating reservoir-based averaging or algorithm-based forecasting regimes.

Summary

The gist: Timescale matching between noise correlation time, reservoir memory, and forecast horizon determines whether correlated noise is filtered or predicted by a physical reservoir system.

Introduction to Physical Reservoir Computing (RC)

Experimental Setup and Signal Generation

The study utilized a thin oxide film-based physical reservoir with defects that form a solid-state memristor with volatile switching dynamics as an experimental platform <ref:2610.10037#pg4>. The synthetic signal was created by mixing a deterministic signal component of two harmonics with added coloured noise <ref:2610.10037#pg4>. The noise correlation times were chosen as τ0 to 1000τ0, where τ0 = 6.3×10−7 s, and the frequency f was chosen to allow 30 periods during the network training and inference <ref:2610.10037#pg6>. The total signal was formed by combining the deterministic signal with the coloured noise μ(t), converting it to a voltage waveform <ref:2610.10037#pg6>.

Performance Analysis via Timescale Matching

The research established two regimes for predictions of noisy signals: reservoir-based averaging and algorithm-based prediction <ref:2610.10037#pg15>. If noise correlation times are well below both the MH time range and the prediction window size, a reservoir-based averaging is taking place, meaning worse overall performance for prediction in comparison to simply reconstructing the deterministic signal <ref:2610.10037#pg15>. Conversely, if the coloured noise correlation time is longer than both the MH time and the prediction window size, then the physical reservoir treats the coloured noise component as part of the signal, allowing reconstruction and prediction of the signal combined with injected noise <ref:2610.10037#pg15>.

Forecasting Regime Index (FRI)

An empirically introduced forecasting regime index (FRI) is used to distinguish between these two regimes by calculating FRI=log10[τ/√τmτw] <ref:2610.10037#pg15>. When FRI < 0, the noise evolves faster (slower) than both the prediction and the MH timescale, indicating reservoir-based averaging or algorithm-based forecasting, respectively <ref:2610.10037#pg15>. The results show that for reconstruction and prediction of noise-injected signals, NRMSE values were at the lowest for the longest noise correlation times <ref:2610.10037#pg9>.

Real-World Data Application

The system was tested with real-world noisy time series, specifically the time-averaged volatility for prices of four cryptocurrencies: Bitcoin (BTC), Ethereum (ETH), Solana (SOL) and Dogecoin (DOGE) <ref:2610.10037#pg10>. The results showed that the NRMSE values decrease as the correlation time increases, in agreement with previous results on forecasting the synthetic noise-injected signal for longer correlation times <ref:2610.10037#pg11>. The lowest NRMSE values corresponded to future window sizes smaller than the past window size of 15.87τ0 but also for correlation times larger than the future window size <ref:2610.10037#pg11>.

Conclusion and Design Principles

The main conceptual advance is that coloured noise is treated not simply as a nuisance, but as a structured component of the learning problem <ref:2610.10037#pg2>. The fading memory of the reservoir defines a practical operating window for noisy temporal inference, suggesting that successful prediction is governed by timescale matching between the input and the hardware <ref:2610.10037#pg3>. This implies that optimal encoding should be determined by autocorrelation structure and spectral content rather than just raw sampling interval <ref:2610.10037#pg19>. The reservoir outputs may also contain richer information than direct regression due to a mixture of timescales of input data with time dynamics and filtering of the RC <ref:2610.10037#pg19>.

Experimental Details

The physical reservoir utilized a nanoporous N2-doped, 50 nm thick niobium oxide layer sandwiched between nanoporous platinum and titanium adhesion layers <ref:2610.10037#pg2>. The noise signal was generated using an exponential autocorrelation function where 〈μ(t1)μ(t2)〉 = D ∙ Var(τ) exp(-t1 - t2/τ), with τ0 = 6.3×10−7 s and D ranging from 0.04 to 40 <ref:2610.10037#pg2>. The analysis involved calculating Memory Capacity (MC) and Nonlinearity (NL), where MC increased up to a maximum of 4.11 at a window size of 15.87τ0 time steps <ref:2610.10037#pg3>. The paper demonstrated that the noise correlation time obtained from predicted waveforms yielded increased correlation time values when the prediction window size was larger than 15.87τ0 <ref:2610.10037#pg4>.

Cryptocurrency Volatility Results

Forecasting the volatility of cryptocurrencies showed that NRMSE values decrease as the correlation time increases, in agreement with previous results on forecasting the synthetic noise-injected signal for longer correlation times <ref:2610.10037#pg11>. The lowest NRMSE values corresponded to future window sizes smaller than the past window size of 15.87τ0 but also for correlation times larger than the future window size <ref:2610.10037#pg4>.

Hardware and Measurement

The device fabrication involved depositing the N2-doped NbOx by magnetron sputtering and patterning it by UV lithography <ref:2610.10037#pg2>. The measurements were conducted using a Keithley 4200ASCS by applying the input signal through the common bottom electrode and taking an electric current measurement out of the four top electrode pads <ref:2610.10037#pg2>. Linear regression was carried out in a manner described in previous work, utilizing Bayesian optimization to help minimize NRMSE <ref:2610.10037#pg20>. The memory capacity (MC) was calculated as MC = ∑r2∞ k=1 [u(n − k), yk(n)] <ref:2610.10037#pg22>. The nonlinearity (NL) was calculated using the formula NL = 1 − r2[ΣijWiju(n), ŷ(n)] <ref:2610.10037#pg22>. The time step comparison showed that ST1 in the supplementary material shows the comparison between time steps for the different measurements discussed in this work and how they relate to τ0=6.3×10−7 s <ref:2610.10037#pg37>. The total number of data points recorded was 18423 points corresponding to the time duration of 3249τ0 <ref:2610.10037#pg21>.

References

Kempelis, A., Narigina, M., Osadcijs, E., Patlins, A. & Romanovs, A. Machine Learning-based Sensor Data Forecasting for Precision Evaluation of Environmental Sensing. in 2023 IEEE 10th Jubilee Workshop on Advances in Information, Electronic and Electrical Engineering (AIEEE) 1–6 (IEEE, Vilnius, Lithuania, 2023). doi:10.1109/AIEEE58915.2023.10135031 <ref:2610.10037#pg23>. Giger, M. L. Machine Learning in Medical Imaging. J. Am. Coll. Radiol. 15, 512–520 (2018) <ref:2610.10037#pg23>. Bochenek, B. & Ustrnul, Z. Machine Learning in Weather Prediction and Climate Analyses—Applications and Perspectives. Atmosphere 13, 180 (2022) <ref:2610.10037#pg24>. Xiao, D. et al. A reduced order model for turbulent flows in the urban environment using machine learning <ref:2610.10037#pg24>. Goldberger, A. L. et al. Fractal dynamics in physiology: Alterations with disease and aging. Proc. Natl. Acad. Sci. 99, 2466–2472 (2002) <ref:2610.10037#pg6>. Bullmore, E. et al. Colored noise and computational inference in neurophysiological (fMRI) time series analysis: Resampling methods in time and wavelet domains. Hum. Brain Mapp. 12, 61–78 (2001) <ref:2610.10037#pg7>. Cui, B., Chen, X., Tang, X., Huang, H. & Liu, X. Robust cubature Kalman filter for GNSS/INS with missing observations and colored measurement noise. ISA Trans. 72, 138–146 (2018) <ref:2610.10037#pg8>. Santamaría‐Gómez, A. & Ray, J. Chameleonic Noise in GPS Position Time Series. J. Geophys. Res. Solid Earth 126, e2020JB019541 (2021) <ref:2610.10037#pg9>. Hasselmann, K. Stochastic climate models: Part I. Theory. Tellus Dyn. Meteorol. Oceanogr. 28, 473 (1976) <ref:2610.10037#pg10>. Lovejoy, S. Using scaling for macroweather forecasting including the pause. Geophys. Res. Lett.

Improvements for AI systems

Improving physical reservoir architectures by employing timescale matching to guide input encoding and architecture design; timescale matching can guide the encoding of input time series and development of physical reservoir architectures that filter, analyse and predict stochastic signal components across distinct temporal scales.
Developing a novel forecasting regime index (FRI) to automatically distinguish between reservoir-based averaging and algorithm-based prediction, allowing the system to dynamically select the appropriate computational mode based on noise characteristics.
Creating a dynamical filter capable of selectively filtering or extracting predictive structure from colored noise based on its correlation time relative to the reservoir's memory horizon and forecast horizon; this allows the system to handle input where noise varying faster than the relevant reservoir memory and forecast horizon is averaged by the reservoir, whereas the temporal structure of slower-varying noise is sufficient for algorithmic forecasting.
Designing an optimal encoding process that maps the informative temporal structure of the signal, including its noise correlations, into the memory timescales where the reservoir is computationally effective, rather than relying solely on raw sampling intervals.
Enhancing volatility prediction accuracy in real-world financial markets by leveraging reservoir architectures to handle complex signals where deterministic and stochastic components cannot be separated explicitly, specifically by exploiting the temporal structure of slower-varying noise.

Abstract

Physical reservoir computing exploits the nonlinear dynamics of physical systems to process time-dependent data with greater energy efficiency than conventional machine learning approaches. However, physical reservoirs have fixed intrinsic response timescales, whereas real-world signals combine deterministic and stochastic components across multiple timescales. Here we show, using a nanoporous niobium oxide reservoir, synthetic noisy signals and cryptocurrency-price volatility, that the relationship among noise correlation time, reservoir memory and forecast horizon determines whether correlated noise is filtered or predicted. Noise varying faster than the relevant reservoir memory and forecast horizon is averaged by the reservoir, whereas the temporal structure of slower-varying noise is sufficient for algorithmic forecasting. We introduce the reservoir memory horizon and forecasting regime index to distinguish these operating regimes. These contributions demonstrate that timescale matching can guide the encoding of input time series and development of physical reservoir architectures that filter, analyse and predict stochastic signal components across distinct temporal scales.